Back

The link prediction problem for social networks

David Liben-NowellJon M. Kleinberg

article2003CIKM2,962 citations

Formalizes the link prediction problem for social networks and systematically evaluates a wide range of graph-theoretic proximity measures on large collaboration datasets, demonstrating that network topology alone holds substantial power to predict future interactions.

Listen

The link prediction problem asks how well future interactions in a social network can be inferred from a single snapshot of its current structure. This question matters because networks of collaborations, organizational ties, or security-relevant contacts evolve rapidly, and organizations increasingly seek to anticipate or encourage new connections using only the observable pattern of existing ones.

The paper set out to formalize this task and test whether a range of proximity measures drawn from graph theory and social-network analysis could extract useful predictive signal from topology alone. The authors constructed five large co-authorship networks from arXiv physics sections, using three-year windows (1994–1996 for training, 1997–1999 for testing) and restricting attention to a core set of active authors. They evaluated roughly twenty predictors—ranging from simple counts of common neighbors and shortest-path distance to more elaborate methods such as Katz summation over all paths, rooted PageRank, SimRank, and low-rank matrix approximations—by measuring how far each method improved upon a random baseline that merely guessed among non-edges in the training graph.

Several predictors substantially outperformed chance, with the best variants of the Katz measure and certain low-rank approximations achieving roughly 40- to 50-fold gains on three of the five datasets. Common-neighbor counting and the Adamic/Adar refinement also performed strongly and consistently, while shortest-path distance proved weaker once the “small-world” noise of distant cross-field links was taken into account. Even when the evaluation was restricted to pairs at graph distance three or greater, most methods still beat random by factors of five to thirty. Overlap among the top predictors was high, yet no single technique dominated every dataset.

These results indicate that network topology encodes substantial latent information about future collaborations, independent of external attributes such as geography or research topic. The practical implication is that organizations could use lightweight topological scores to surface promising but unrealized partnerships, and that security or intelligence applications could generate plausible hypotheses about unobserved ties. The findings also suggest that models of network growth should be judged not only by their ability to reproduce global statistics but by their success at local link-prediction tasks.

Further gains are likely possible. Promising directions include weighting recent edges more heavily, exploiting the richer bipartite author-paper representation, tuning parameters on temporal slices of the training data, and adapting support-estimation techniques from machine learning. Efficiency improvements for very large graphs remain necessary before the methods can be deployed at scale.

The study is limited to co-authorship data in physics, relies on noisy name-based author disambiguation, and treats all training edges equally; performance may therefore differ in other domains or with richer node attributes. Even the strongest predictor reached only about 16 percent accuracy, confirming that many new links arise for reasons outside the observed network.

Cornell UniversityMassachusetts Institute of Technologycoauthorshiplink predictionnetwork architectureLink PredictionSocial NetworkProximity Measure
Liben-Nowell et al (2003).pdf
  • Paper: The PageRank Citation Ranking : Bringing Order to the Web, Lawrence Page et al. (1999). Understanding PageRank's foundational link-analysis algorithm provides the essential algorithmic basis for the network proximity measures adapted in the source paper.
  • Paper: Mining the network value of customers, Pedro M. Domingos et al. (2001). This early work on valuing customer network ties provides foundational context for modeling interactions and influence within social networks.
  • Paper: node2vec: Scalable Feature Learning for Networks, Aditya Grover et al. (2016). This paper extends the link prediction problem by introducing scalable feature learning methods that automatically generate node and edge representations.
  • Paper: LINE: Large-scale Information Network Embedding, Jian Tang et al. (2015). This work builds directly upon network proximity concepts to develop scalable information network embeddings for link prediction.
  • Paper: Variational Graph Auto-Encoders, Thomas N. Kipf et al. (2016). This paper extends graph-based link prediction into the realm of variational auto-encoders by combining network structure with node features.
Cover for The link prediction problem for social networks

Abstract

Given a snapshot of a social network, can we infer which new interactions among its members are likely to occur in the near future? We formalize this question as the link prediction problem, and develop approaches to link prediction based on measures for analyzing the “proximity” of nodes in a network. Experiments on large co-authorship networks suggest that information about future interactions can be extracted from network topology alone, and that fairly subtle measures for detecting node proximity can outperform more direct measures.

Topics

coauthorshiplink predictionnetwork architecturenetwork topologytopology

Keywords

Link PredictionSocial NetworkProximity MeasureCommon NeighborAdamic/AdarKatzRooted PageRankLow‑Rank ApproximationUnseen BigramSmall World

Table of Contents

  • 1 Introduction
  • 2 Data and Experimental Setup
  • 3 Methods for Link Prediction
  • 4 Results and Discussion
  • References

Knowls

  1. Knowl 1 — Link Prediction Problem Formulation and Evaluation Protocol

    definition

    The link prediction problem formalizes the task of inferring future interactions in a social network using only the network's topological snapshot.

    Let G=(V,E)G = (V, E) be a temporal social network where each edge e=⟨u,v⟩∈Ee = \langle u, v \rangle \in E denotes an interaction at timestamp t(e)t(e). For a time interval [t,t′][t, t'], let G[t,t′]G[t, t'] denote the subgraph of edges with timestamps within [t,t′][t, t']. Given a training interval [t0,t0′][t_0, t_0'] and a test interval [t1,t1′][t_1, t_1'] with t0<t0′<t1<t1′t_0 < t_0' < t_1 < t_1', the input is the training collaboration graph Gcollab:=G[t0,t0′]=(A,Eold)G_{\text{collab}} := G[t_0, t_0'] = (A, E_{\text{old}}). The set of true future collaborations to predict is: Enew:={⟨u,v⟩∣u,v∈A,⟨u,v⟩∈G[t1,t1′],⟨u,v⟩∉Eold}E_{\text{new}} := \{\langle u, v \rangle \mid u, v \in A, \langle u, v \rangle \in G[t_1, t_1'], \langle u, v \rangle \notin E_{\text{old}}\}

    To ensure evaluated nodes are active during both periods, predictions are evaluated on the subset Core⊆A\text{Core} \subseteq A of nodes that have degree at least κtraining\kappa_{\text{training}} in G[t0,t0′]G[t_0, t_0'] and degree at least κtest\kappa_{\text{test}} in G[t1,t1′]G[t_1, t_1'] (with κtraining=κtest=3\kappa_{\text{training}} = \kappa_{\text{test}} = 3). The evaluation target is Enew∗:=Enew∩(Core×Core)E^*_{\text{new}} := E_{\text{new}} \cap (\text{Core} \times \text{Core}), with n:=∣Enew∗∣n := |E^*_{\text{new}}|.

    A link predictor outputs a ranked list LpL_p of candidate pairs in (A×A)∖Eold(A \times A) \setminus E_{\text{old}} in descending order of a topological score score(x,y)\text{score}(x, y). The predictor's performance is measured by the size of the intersection between the top nn pairs in LpL_p restricted to Core×Core\text{Core} \times \text{Core} and the target set Enew∗E^*_{\text{new}}, expressed as a factor of improvement over a baseline random predictor that selects pairs uniformly at random from (A×A)∖Eold(A \times A) \setminus E_{\text{old}}.

  2. Knowl 2 — Neighborhood-Based Proximity Measures for Link Prediction

    model/method

    Neighborhood-based link predictors assign a connection score score(x,y)\text{score}(x, y) to node pairs ⟨x,y⟩\langle x, y \rangle based exclusively on the local neighborhood structure in the training graph GcollabG_{\text{collab}}, where Γ(x)\Gamma(x) denotes the set of neighbors of node xx.

    1. Common Neighbors: Measures the raw overlap of node neighborhoods: score(x,y):=∣Γ(x)∩Γ(y)∣\text{score}(x, y) := |\Gamma(x) \cap \Gamma(y)|

    2. Jaccard's Coefficient: Computes the normalized intersection of neighbors: score(x,y):=∣Γ(x)∩Γ(y)∣∣Γ(x)∪Γ(y)∣\text{score}(x, y) := \frac{|\Gamma(x) \cap \Gamma(y)|}{|\Gamma(x) \cup \Gamma(y)|}

    3. Adamic/Adar: Refines neighbor counting by weighting shared neighbors inversely by their log-degree, thereby penalizing shared high-degree hubs: score(x,y):=∑z∈Γ(x)∩Γ(y)1log⁡∣Γ(z)∣\text{score}(x, y) := \sum_{z \in \Gamma(x) \cap \Gamma(y)} \frac{1}{\log |\Gamma(z)|}

    4. Preferential Attachment: Scores pairs based on the product of their degrees, following the principle that rich nodes preferentially acquire new links: score(x,y):=∣Γ(x)∣⋅∣Γ(y)∣\text{score}(x, y) := |\Gamma(x)| \cdot |\Gamma(y)|

    Candidate pairs are ranked in descending order of score(x,y)\text{score}(x, y) to predict new links.

  3. Knowl 3 — Path-Based and Random-Walk Proximity Measures for Link Prediction

    model/method

    Path-based and random-walk methods score candidate node pairs ⟨x,y⟩\langle x, y \rangle by implicitly aggregating the ensemble of all paths connecting them in the training graph GcollabG_{\text{collab}}:

    1. Katz Measure: Directly sums all paths between xx and yy, exponentially damped by path length ℓ\ell using an attenuation parameter β>0\beta > 0: score(x,y):=∑ℓ=1∞βℓ⋅∣pathsx,y⟨ℓ⟩∣=[(I−βM)−1−I]x,y\text{score}(x, y) := \sum_{\ell=1}^\infty \beta^\ell \cdot |\text{paths}_{x,y}^{\langle \ell \rangle}| = \left[(I - \beta M)^{-1} - I\right]_{x,y} where MM is the adjacency matrix of GcollabG_{\text{collab}}. In the unweighted variant, Mu,v∈{0,1}M_{u,v} \in \{0, 1\}; in the weighted variant, Mu,vM_{u,v} equals the number of training interactions between uu and vv.

    2. Hitting Time and Commute Time: Let Hx,yH_{x,y} be the expected number of steps for a simple random walk on GcollabG_{\text{collab}} starting at xx to first reach yy. The commute time is Cx,y:=Hx,y+Hy,xC_{x,y} := H_{x,y} + H_{y,x}. Raw proximity scores are −Hx,y-H_{x,y} and −Cx,y-C_{x,y}. To counteract the bias where Hx,yH_{x,y} is small merely because yy has a high stationary distribution weight πy\pi_y, stationary-normed variants are defined as: score(x,y):=−Hx,y⋅πyorscore(x,y):=−(Hx,y⋅πy+Hy,x⋅πx)\text{score}(x, y) := -H_{x,y} \cdot \pi_y \quad \text{or} \quad \text{score}(x, y) := -(H_{x,y} \cdot \pi_y + H_{y,x} \cdot \pi_x)

    3. Rooted PageRank: Defines score(x,y)\text{score}(x, y) as the stationary probability of yy in a random walk with restart that, at each step, resets to xx with probability α∈(0,1)\alpha \in (0, 1) and transitions to a uniformly random neighbor of the current node with probability 1−α1 - \alpha.

    4. SimRank: Defines structural similarity recursively such that similarity(x,x):=1\text{similarity}(x, x) := 1 and for x≠yx \neq y: similarity(x,y):=γ⋅∑a∈Γ(x)∑b∈Γ(y)similarity(a,b)∣Γ(x)∣⋅∣Γ(y)∣\text{similarity}(x, y) := \gamma \cdot \frac{\sum_{a \in \Gamma(x)} \sum_{b \in \Gamma(y)} \text{similarity}(a, b)}{|\Gamma(x)| \cdot |\Gamma(y)|} for a decay parameter γ∈[0,1]\gamma \in [0, 1]. This corresponds to E[γℓ]\mathbb{E}[\gamma^\ell], where ℓ\ell is the random variable representing the step at which independent random walks from xx and yy first meet.

  4. Knowl 4 — Meta-Approaches for Link Predictor Enhancement

    model/method

    Meta-approaches modify the representation of the training graph or proximity scores to improve link prediction accuracy across base predictors:

    1. Low-Rank Approximation (Singular Value Decomposition): Given the adjacency matrix MM of GcollabG_{\text{collab}}, compute its rank-kk truncated SVD approximation MkM_k. Proximity is scored via:
    • The direct matrix entry: score(x,y):=(Mk)x,y\text{score}(x, y) := (M_k)_{x,y}
    • The inner product of reconstructed rows: score(x,y):=⟨rk(x),rk(y)⟩=(MkMkT)x,y\text{score}(x, y) := \langle r_k(x), r_k(y) \rangle = (M_k M_k^T)_{x,y}
    • Truncated Katz: evaluating [(I−βMk)−1−I]x,y[(I - \beta M_k)^{-1} - I]_{x,y}

    1. Unseen Bigrams (Collaborative Similarity Smoothing): In analogy to estimating frequencies of unseen word pairs in language modeling, node proximity estimates are smoothed using the set Sx⟨δ⟩S_x^{\langle \delta \rangle} of the δ\delta most similar nodes to xx under a base metric score(x,⋅)\text{score}(x, \cdot): scoreunweighted∗(x,y):=∣{z∣z∈Γ(y)∩Sx⟨δ⟩}∣\text{score}^*_{\text{unweighted}}(x, y) := |\{z \mid z \in \Gamma(y) \cap S_x^{\langle \delta \rangle}\}| scoreweighted∗(x,y):=∑z∈Γ(y)∩Sx⟨δ⟩score(x,z)\text{score}^*_{\text{weighted}}(x, y) := \sum_{z \in \Gamma(y) \cap S_x^{\langle \delta \rangle}} \text{score}(x, z)

    2. Clustering / Edge Pruning: Low-confidence edges in the training graph EoldE_{\text{old}} are pruned by computing score(u,v)\text{score}(u, v) for all existing edges ⟨u,v⟩∈Eold\langle u, v \rangle \in E_{\text{old}} and deleting the (1−ρ)(1 - \rho) fraction with the lowest scores. The base predictor is then recomputed on the cleaned subgraph.

  5. Knowl 5 — Comparative Performance of Topological Link Predictors on arXiv Networks

    data/table

    Topological link prediction methods were evaluated on five arXiv physics collaboration networks (astro-ph, cond-mat, gr-qc, hep-ph, hep-th) using training data from 1994–1996 and test data from 1997–1999. Predictors were assessed by their factor improvement over a random prediction baseline when selecting the top n=∣Enew∗∣n = |E^*_{\text{new}}| candidates among active Core authors.

    Predictor astro-ph cond-mat gr-qc hep-ph hep-th
    Random baseline correct probability 0.475% 0.147% 0.341% 0.207% 0.153%
    Graph distance (shortest path) 9.6 25.3 21.4 12.2 29.2
    Common neighbors 18.0 41.1 27.2 27.0 47.2
    Preferential attachment 4.7 6.1 7.6 15.2 7.5
    Adamic/Adar 16.8 54.8 30.1 33.3 50.5
    Jaccard's coefficient 16.4 42.3 19.9 27.7 41.7
    SimRank (γ=0.8\gamma = 0.8) 14.6 39.3 22.8 26.1 41.7
    Rooted PageRank (α=0.15\alpha = 0.15) 16.6 41.1 27.2 27.6 42.6
    Katz weighted (β=0.005\beta = 0.005) 13.4 54.8 30.1 24.0 52.2
    Katz unweighted (β=0.005\beta = 0.005) 16.8 41.7 37.5 24.2 49.7
    Low-rank inner product (rank=256\text{rank} = 256) 14.6 47.1 29.4 32.4 47.2
    Unseen bigrams weighted Katz (δ=8\delta = 8) 16.9 38.1 25.0 24.2 51.3
    Katz clustering (ρ=0.15\rho = 0.15) 12.0 46.5 47.1 21.1 44.2

    Topological proximity contains strong latent signals for future collaborations, outperforming random chance by factors of 40 to 50 on multiple datasets. While simple heuristics like common neighbors and Adamic/Adar perform robustly, Katz centrality and its variants (clustering and low-rank approximation) achieve the highest overall predictive performance across datasets.

  6. Knowl 6 — The Small-World Limitation and the Distance-3 Prediction Task

    empirical result

    In co-authorship social networks, between 71% (hep-ph: 2,349 of 3,294) and 83% (cond-mat: 960 of 1,150; gr-qc: 332 of 400) of newly formed collaborations appear between author pairs at shortest-path distance 3 or greater in the training graph GcollabG_{\text{collab}}.

    Because social networks exhibit the small-world property, almost all nodes are connected via very short paths, meaning shortest-path distance alone cannot discriminate well among candidate pairs at distance 2. Furthermore, common neighbor and direct neighborhood overlap measures assign a score of zero to all pairs at distance ≥3\ge 3.

    On the "Distance-3 Task"—where all distance-2 candidate pairs are removed and predictors must identify new links strictly among pairs at initial distance ≥3\ge 3—random-walk and all-path ensemble predictors significantly outperform baseline distance predictors (which achieve only 3.1x to 8.4x over random):

    • Unweighted Katz (β=0.005\beta = 0.005) achieves factor improvements over random of 9.2x (astro-ph), 11.9x (cond-mat), 30.6x (gr-qc), 5.1x (hep-ph), and 18.0x (hep-th).
    • Rooted PageRank (α=0.15\alpha = 0.15) achieves 5.4x (astro-ph), 11.9x (cond-mat), 18.0x (gr-qc), 11.1x (hep-ph), and 20.0x (hep-th).
    • SimRank (γ=0.8\gamma = 0.8) achieves 6.0x (astro-ph), 14.4x (cond-mat), 10.6x (gr-qc), 7.7x (hep-ph), and 22.0x (hep-th).
    • Low-rank matrix approximations achieve factor improvements as high as 32.8x over random on gr-qc.

  7. Knowl 7 — Impact of Topical Breadth on Network Link Predictability

    empirical result

    The relative predictive accuracy of topological link prediction methods (measured as factor improvement over random guessing) scales directly with the topical breadth and diversity of the underlying network dataset.

    Evaluating the common neighbors predictor against a random baseline across datasets of varying scope yields:

    • Narrow focus (STOC and FOCS theoretical computer science conference proceedings): 6.1x improvement over random.
    • Individual physics subdisciplines (arXiv sections): 18.0x to 47.2x improvement over random.
    • Combined multi-discipline physics (all 5 arXiv sections merged): 71.2x improvement over random.
    • Broad multi-domain academic archive (CiteSeer web crawl across all computer science domains): 147.0x improvement over random.

    In narrow research communities, the boundary between subfields is minimal and interactions are heavily stochastic. In broader, topically diverse networks, topological proximity strongly separates distinct disciplines, allowing topological predictors to vastly outperform random selection.

  8. Knowl 8 — Prediction Overlap Patterns and Network Structural Heterogeneity

    empirical result

    Pairwise overlap analysis of top predicted edges reveals distinct clustering behaviors among link prediction algorithms and structural differences across network domains:

    1. Predictor Clustered Behavior:
    • Katz measure, low-rank inner product, and Adamic/Adar produce highly overlapping sets of candidate edges (e.g., out of 1,150 predictions on cond-mat, Adamic/Adar shares 1,011 predictions with weighted Katz and 905 with low-rank inner product).
    • Rooted PageRank, SimRank, and Jaccard coefficient form a second cluster of mutually similar prediction sets.
    • Hitting time produces predictions largely disjoint from all other predictors (sharing only 87 to 247 predictions with the other methods on cond-mat), despite achieving competitive standalone performance.
    1. Cross-Domain Topological Heterogeneity:
    • In gr-qc, low-rank SVD methods achieve optimal performance at rank k=1k = 1 (44.9x improvement over random), indicating a simpler, low-dimensional organizational structure than the other four sections, which peak at intermediate ranks (k=64k = 64 to 10241024).
    • In hep-ph, preferential attachment (degree product) achieves a 15.2x improvement over random, surpassing the graph distance predictor (12.2x), demonstrating that raw author productivity/degree is unusually informative in high-energy physics phenomenology compared to other disciplines.

Coverage note — None was omitted; all key contributions—including problem formulation, all proximity algorithms, meta-approaches, comprehensive benchmark results on arXiv, distance-3 experiments, topical breadth scaling, and algorithm overlap analysis—are fully covered.

References

  1. 1.Lada A. Adamic and Eytan Adar. Friends and neighbors on the web. Social Networks, 25(3):211–230, July 2003.
  2. 2.A. L. Barabasi, H. Jeong, Z. Néda, E. Ravasz, A. Schubert, and T. Vicsek. Evolution of the social network of scientific collaboration. Physica A, 311(3–4):590–614, 2002.
  3. 3.Sergey Brin and Lawrence Page. The anatomy of a large-scale hypertextual Web search engine. Computer Networks and ISDN Systems, 30(1–7):107–117, 1998.
  4. 4.Rodrigo De Castro and Jerrold W. Grossman. Famous trails to Paul Erdös. Mathematical Intelligencer, 21(3):51–63, 1999.
  5. 5.Jörn Davidsen, Holger Ebel, and Stefan Bornholdt. Emergence of a small world from local interactions: Modeling acquaintance networks. Physical Review Letters, 88(128701), 2002.
  6. 6.Scott Deerwester, Susan T. Dumais, George W. Furnas, Thomas K. Landauer, and Richard Harshman. Indexing by latent semantic analysis. Journal of the American Society for Information Science, 41(6):391–407, 1990.
  7. 7.Ute Essen and Volker Steinbiss. Cooccurrence smoothing for stochastic language modeling. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 1, pages 161–164, 1992.
  8. 8.Debra S. Goldberg and Frederick P. Roth. Assessing experimentally derived interactions in a small world. In Proceedings of the National Academy of Sciences USA, volume 100, pages 4372–4376, April 2003.
  9. 9.Jerrold W. Grossman. The evolution of the mathematical research collaboration graph. In Proceedings of the Southeast Conference on Combinatorics, Graph Theory, and Computing, March 2002.
  10. 10.Glen Jeh and Jennifer Widom. SimRank: A measure of structural-context similarity. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, July 2002.
  11. 11.Emily M. Jin, Michelle Girvan, and M. E. J. Newman. The structure of growing social networks. Physical Review Letters E, 64(046132), 2001.
  12. 12.Leo Katz. A new status index derived from sociometric analysis. Psychometrika, 18(1):39–43, March 1953.
  13. 13.H. Kautz, B. Selman, and M. Shah. ReferralWeb: Combining social networks and collaborative filtering. Communications of the ACM, 30(3), March 1997.
  14. 14.Valdis Krebs. Mapping networks of terrorist cells. Connections, 24(3):43–52, Winter 2002.
  15. 15.Lillian Lee. Measures of distributional similarity. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, pages 25–32, 1999.
  16. 16.Michael Mitzenmacher. A brief history of lognormal and power law distributions. In Proceedings of the Allerton Conference on Communication, Control, and Computing, pages 182–191, 2001.
  17. 17.M. E. J. Newman. Clustering and preferential attachment in growing networks. Physical Review Letters E, 64(025102), 2001.
  18. 18.M. E. J. Newman. The structure of scientific collaboration networks. Proceedings of the National Academy of Sciences USA, 98:404–409, 2001.
  19. 19.M. E. J. Newman. The structure and function of networks. Computer Physics Communications, 147:40–45, 2002.
  20. 20.M. E. J. Newman. The structure and function of complex networks. SIAM Review, 45:167–256, 2003.
  21. 21.Christopher Palmer, Phillip Gibbons, and Christos Faloutsos. ANF: A fast and scalable tool for data mining in massive graphs. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Jul 2002.
  22. 22.A. Popescul and L. Ungar. Statistical relational learning for link prediction. In Workshop on Learning Statistical Models from Relational Data at the International Joint Conference on Artificial Intelligence, 2003.
  23. 23.P. Raghavan. Social networks: From the web to the enterprise. IEEE Internet Computing, 6(1):91–94, January/February 2002.
  24. 24.Gerard Salton and Michael J. McGill. Introduction to Modern Information Retrieval. McGraw-Hill, 1983.
  25. 25.Bernhard Schölkopf, John C. Platt, John Shawe-Taylor, Alex J. Smola, and Robert C. Williamson. Estimating the support of a high-dimensional distribution. Technical Report MSR-TR-99-87, Microsoft Research, 1999.
  26. 26.Ben Taskar, Ming-Fai Wong, Pieter Abbeel, and Daphne Koller. Link prediction in relational data. In Proceedings of Neural Information Processing Systems, 2004. To appear.
  27. 27.Duncan J. Watts. Small Worlds. Princeton University Press, 1999.
  28. 28.Duncan J. Watts and Steven H. Strogatz. Collective dynamics of ‘small-world’ networks. Nature, 393:440–442, 1998.

Citation

MLA
Liben-Nowell, D., and J. Kleinberg. “The Link Prediction Problem for Social Networks”. Proceedings of the Twelfth International Conference on Information and Knowledge Management, 2003, pp. 556–59, https://doi.org/10.1145/956863.956972.
APA
Liben-Nowell, D., & Kleinberg, J. (2003). The link prediction problem for social networks. Proceedings of the Twelfth International Conference on Information and Knowledge Management, 556–559. https://doi.org/10.1145/956863.956972
Chicago
Liben-Nowell, D., and J. Kleinberg. 2003. “The Link Prediction Problem for Social Networks”. Proceedings of the Twelfth International Conference on Information and Knowledge Management, 556–59. https://doi.org/10.1145/956863.956972.
Harvard
Liben-Nowell, D. and Kleinberg, J. (2003) “The link prediction problem for social networks”, Proceedings of the twelfth international conference on Information and knowledge management. ACM, pp. 556–559. Available at: https://doi.org/10.1145/956863.956972.
Vancouver
1. Liben-Nowell D, Kleinberg J (2003) The link prediction problem for social networks. In: Proceedings of the twelfth international conference on Information and knowledge management. ACM, pp 556–559

BibTeX

@inproceedings{Liben_Nowell_2003, series={CIKM03}, title={The link prediction problem for social networks}, url={http://dx.doi.org/10.1145/956863.956972}, DOI={10.1145/956863.956972}, booktitle={Proceedings of the twelfth international conference on Information and knowledge management}, publisher={ACM}, author={Liben-Nowell, David and Kleinberg, Jon}, year={2003}, month=Nov, pages={556–559}, collection={CIKM03} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF